Aspect Based sentiment analysis (ABSA) is concerned with mining opinions from text about specific entities and their aspects. Nowadays, people easily spill their opinions on the internet about anything. In this thesis, we focused on customer reviews of smartphones to do ABSA. This could be done by using supervised machine learning, however annotating is very time con- suming. A lot of web shops summarise reviews by listing positive and negative points above the reviews. So in fact, the annotation has already been done. In this thesis, we researched whether distant supervised ma- chine learning can achieve comparable results as with supervised machine learning. As data, 26,565 customer reviews of smartphones from the web site pdashop.nl were used. This data consisted of 94,307 distant supervised annotated points. Twelve categories were defined based on the aspects of a smartphone. For each category a word list was created to match sentences of a review to a category. Using a Support Vector Machine we created for each category one model. As features we used POS tags, different types of N-grams, word frequencies and TF-IDF-scores. 887 sentences annotated by two annotators, divided over the twelve categories were used to evaluate the system. Dependent on the category, achieved accuracies were between 0.22 and 0.72. In this research we didn’t achieve results comparable to su- pervised machine learning (0.79). Highly skewed data to positive making it hard to detect negative sentiments and not having the ability to train on neutral sentiments were the main causes for low accuracies.

The complete Bachelor Thesis can be downloaded here:
Thesis